Automation & Workflows · Part 3 of 5: The Digital Employee Playbook
How Do You Turn Judgment Calls Into Rules an AI Can Follow?
To turn a judgment call into a rule an AI can follow, ask whoever makes the decision why, then write the answer as a condition, an action, and an exception. Most "it depends" answers come down to a few checkable facts: a date, a dollar amount, a document, a yes or no. Those become rules. Decisions that do not reduce to facts are where a person stays in the loop.
This is part three of the Digital Employee Playbook. Part two documented the steps of a workflow and flagged every "it depends." This part resolves them.
Why do judgment calls block automation?
Because the person making them cannot see the rule anymore. An experienced referral coordinator glances at a packet and knows it is incomplete. Ask why, and the first answer is usually "you can just tell." Underneath, there are always specifics: the clinical notes are missing, the order is unsigned, the insurance card is the back side only.
An AI system needs those specifics written down. Without them it will guess, and a guess is the one thing you cannot audit.
How do you find the rule behind a decision?
Interview the person who makes it, using real cases instead of hypotheticals. Four questions do most of the work:
- "Walk me through the last three times you made this call." Real cases surface the facts people actually check.
- "What do you look at first?" The first thing checked is usually the deciding factor.
- "What would make you decide the opposite way?" This finds the exceptions.
- "Who do you ask when you are not sure?" This finds where the rule runs out and a person takes over.
Write down every fact mentioned. Most decisions turn out to depend on two or three of them.
How do you write decision rules an AI can follow?
Use a decision table: one row per situation, with the facts on the left and the action on the right. Here is step 4 from the referral workflow in part two, the insurance check, written as a table:
| Eligibility result | Plan accepted? | Auth required? | Action |
|---|---|---|---|
| Active | Yes | No | Schedule the patient |
| Active | Yes | Yes | Create an authorization task for staff; schedule after approval |
| Active | No | Any | Offer the self-pay estimate; send to billing if the patient wants to proceed |
| Inactive | Any | Any | Ask the patient for current insurance; recheck when received |
| Not found | Any | Any | Ask the referring office to confirm the member ID; recheck |
Every combination has an action, so nothing falls through. Two writing rules keep tables usable:
- Use facts a system can check. "Plan accepted" works because the practice keeps a list of accepted plans. "Good insurance" does not.
- Give every row an owner. If the action is "staff," name the role, so the handoff has somewhere to go.
The same method works outside healthcare. A personal injury firm's intake decision often turns on a short list of facts: when the incident happened, whether there was an injury, whether the person already has a lawyer, and what kind of case it is. Each of those can be asked and checked, so each can be a rule, while the decision to take a borderline case stays with an attorney.
Where should AI use judgment, and where should it follow rules?
AI is strongest at reading. It can pull a patient's name, member ID, and reason for referral out of a blurry fax, sort an email into "new request" or "follow-up," or summarize a voicemail. Those are judgment tasks that used to need a person, and language models do them well.
Deciding what happens next is different. When the consequence matters, a written rule should make that call, because a rule can be checked, explained, and changed. The pattern that works in the AI workflows Cream Digital builds is: the AI reads and classifies, the rules decide, and people handle whatever the rules cannot. That pattern is also what separates an AI workflow from an AI agent.
How do you test decision rules before trusting them?
Run them on the past before you run them on the future.
- Replay last month. Take 50 to 100 real cases and run them through the rules on paper or in a test system.
- Compare with what staff did. Every disagreement is either a missing rule or a case where staff were inconsistent. Both are worth knowing.
- Run in shadow mode. For the first weeks live, the system proposes the decision and a person approves it. Track how often the person changes it.
- Promote rows, not the whole table. When a row is approved unchanged week after week, let it run on its own. Keep the rest under review.
Some decisions will never reduce to rules, and that is fine. Deciding which ones, and what to do with them, is part four of this series.
Key facts
- A decision rule has three parts: the condition that is checked, the action taken, and the exception that overrides it.Source: Cream Digital, Digital Employee Playbook
- Running proposed rules against last month's real cases, before going live, shows where the rules and staff decisions disagree.Source: Cream Digital, Digital Employee Playbook
- In a well-designed AI workflow, the AI reads and classifies messy inputs, written rules decide the next step, and people handle what the rules cannot.Source: creamdigital.ai/blog/what-is-an-ai-workflow
Frequently asked questions
Can an AI system make judgment calls on its own?
Language models can weigh unstructured information and make a reasonable call, but a business usually cannot audit why. For decisions with real consequences, it is safer to have the AI gather and summarize the facts and let written rules or a person make the call.
What if my staff make the same decision differently?
That is common, and it is one of the most useful things documentation reveals. Pick the rule the business actually wants, write it down, and train both the staff and the AI on it. Consistency is often worth as much as the automation.
How many rules does a typical workflow need?
Fewer than people expect. Most workflows have three to six real decision points, and most decisions turn on two or three facts. The work is in finding them, not in the number of rules.